Vortex5

/

Prototype-X-12b

like 12

Hi,

this is the first really good merge of KansenSakura models in my experience. I'm using Q4_K_S quant (with imatrix) with quantized KV cache (q4_0), so you could already say logical and formatting issues are expected, but not in this case. The model reads character cards well and there's zero formatting issues yet (after 2 hours). The model itself is smart and creative, which was expected. Only the formatting was a deal breaking for me in the original models. Hope this helps.

Thank you for your work, and thanks to Retreatcost for the base models as well.